Calibration des plateformes

Calibration compares the probability the market assigned to each outcome ~24h before resolution against the realised result, over resolved markets with provider-confirmed outcomes and ONE captured price snapshot in the inclusive 20-28h window before resolution. Select the retained timeline capture nearest 24h; equal-distance ties use the earlier capture. Validate that selected book without falling back to another capture if it is invalid or partial. The separate winner-only t24hProbability is a cohort gate, not the scored book. Downsampling can remove an otherwise available window capture, so historical coverage is limited to retained observations. That is a single pre-resolution observation, NOT a claim of continuous 24h history: a market with no snapshot that far back (most same-day markets) is excluded entirely rather than scored on a nearer capture. Reliability bins group (outcome, probability) pairs; a well-calibrated venue's realised win rate tracks its predicted probability. calibrationError is the event-weighted mean gap (ECE, lower is better): each resolved event contributes one unit regardless of outcome count, preventing large outcome catalogs from dominating venue comparisons. Plausibly complete books are normalized to remove overround. An event is scored only when the t-24h snapshot captured the market's COMPLETE outcome book: a partial snapshot would make inclusion depend on whether the eventual winner happened to be captured, which biases the result. Markets with no capture in that window and partial-book snapshots are excluded. Invalid books, including missing or nonnumeric prices, are excluded; a genuine numeric zero remains valid. We store ONE resolved winner per event, so an event is scored only when its recorded structure makes that winner the truth for every outcome: a named Yes/No pair, a single provider market with two sides, or a provider-declared single-winner book (for example Polymarket negRisk or Kalshi mutually_exclusive). Non-exclusive books (threshold, spread and total ladders, price-hit ladders and independent baskets) can resolve several markets YES, and every YES market other than the stored winner would be scored as a loss, so they are excluded and counted as nonExclusiveBook; exclusive books without a recorded single-winner flag, and books with a single stored outcome, are excluded the same way. Exclusivity is never inferred from prices. A single-winner book is scored only when its prices sum to at least 95, the same floor that marks a plausibly complete book for overround removal. Below it, probability mass is not assigned to any outcome: on Kalshi a stored 0 is almost always a contract that had never traded, not a 0% price, and in such books the eventual winner was often one of those zeros. Those books cannot be normalized honestly, so they are excluded and counted as unpricedMassBook. The per-venue counts are published in `excluded` on scored and pending rows; the exclusion rate differs sharply by venue, so compare sampleSize alongside calibrationError.

The final-price series scores the probability the venue itself reported as the market's last pre-resolution price (persisted at ingest) against the realised result, over resolved two-outcome markets whose winner is one of the outcomes and whose both outcome prices are present. Lead time before resolution is unknown and not uniform, so these figures are NOT comparable with the t-24h series above and are reported as a separate basis. Plausibly complete books (price sum 95-125) are normalized to remove overround; others are read raw. The same single-winner rule as the t-24h series applies: a two-outcome market is scored only when it is a named Yes/No pair, one provider market, or a provider-declared single-winner book, and only when its two prices sum to at least 95. Currently published only for venues whose persisted resolved prices were verified to be genuine trading prices rather than settled 0/100 values (Futuur, verified 2026-08-18 across a monotone 2016-2026 reliability curve). Per-year rows below the minimum sample are withheld.

The own-capture series scores the LAST probability CoinRithm itself captured strictly before the provider-reported resolution time (from the durable frozen timeline) against the realised result. Lead time varies and is typically SHORT — per-venue median and p90 lead seconds are published — so with a median lead of minutes this is a near-final-price integrity check over our own observations, NOT a forecast-skill measure, and is not comparable with the t-24h series. An event is scored only when the captured point holds the market's complete outcome book (the same complete-book rule as the t-24h series: a partial capture makes inclusion depend on whether the winner was captured), and only when the book has recorded single-winner structure (the same single-winner rule as the t-24h series; non-exclusive ladders and baskets are excluded) whose prices sum to at least 95 (the same priced-book rule). Plausibly complete books (price sum 95-125) are normalized to remove overround; others are read raw. Events are unit-weighted regardless of outcome count. Coverage grows with forward capture and spans venues the t-24h series undersamples. Per-year rows below the minimum sample are withheld.